PoC Adaptive round robin repartitioning

datafusion/core/src/physical_optimizer/enforce_distribution.rs

@@ -836,7 +836,8 @@ fn add_roundrobin_on_top(
     n_target: usize,
 ) -> Result<DistributionContext> {
     // Adding repartition is helpful:
-    if input.plan.output_partitioning().partition_count() < n_target {
+    // TODO: should we only do for certain sources?
+    if input.plan.output_partitioning().partition_count() < n_target || input.plan.children().is_empty() {
         // When there is an existing ordering, we preserve ordering
         // during repartition. This will be un-done in the future
         // If any of the following conditions is true

datafusion/physical-plan/src/repartition/mod.rs

@@ -779,9 +779,11 @@ impl RepartitionExec {
         metrics: RepartitionMetrics,
         context: Arc<TaskContext>,
     ) -> Result<()> {
+        let is_round_robin = matches!(partitioning, Partitioning::RoundRobinBatch(_));
+        let num_partitions = partitioning.partition_count();
+
         let mut partitioner =
             BatchPartitioner::try_new(partitioning, metrics.repartition_time.clone())?;
-
         // execute the child operator
         let timer = metrics.fetch_time.timer();
         let mut stream = input.execute(partition, context)?;
@@ -801,22 +803,65 @@ impl RepartitionExec {
                 None => break,
             };
 
-            for res in partitioner.partition_iter(batch)? {
+            'next_batch: for res in partitioner.partition_iter(batch)? {
                 let (partition, batch) = res?;
                 let size = batch.get_array_memory_size();
 
-                let timer = metrics.send_time[partition].timer();
-                // if there is still a receiver, send to it
-                if let Some((tx, reservation)) = output_channels.get_mut(&partition) {
-                    reservation.lock().try_grow(size)?;
-
-                    if tx.send(Some(Ok(batch))).await.is_err() {
-                        // If the other end has hung up, it was an early shutdown (e.g. LIMIT)
-                        reservation.lock().shrink(size);
-                        output_channels.remove(&partition);
+                if is_round_robin {
+                    // try out the default (round robin) partition, then all other partitions once starting with partition + 1
+                    for p in std::iter::once(partition).chain(partition+1..num_partitions).chain(0..partition) {
+                        let timer = metrics.send_time[partition].timer();
+
+                        if let Some((tx, reservation)) = output_channels.get_mut(&p) {
+                            reservation.lock().try_grow(size)?;
+
+                            let sent = tx.send(Some(Ok(batch.clone()))).now_or_never();
+
+                            match sent {
+                                Some(res) if res.is_err() => {
+                                    // If the other end has hung up, it was an early shutdown (e.g. LIMIT)
+                                    reservation.lock().shrink(size);
+                                    output_channels.remove(&partition);
+                                }
+                                Some(_) => {
+                                    timer.done();
+                                    continue 'next_batch;
+                                }
+                                None => {
+                                    timer.done();
+                                }
+                            }
+                        }
+                    }
+                    // not yet succeeded to send to any partition, now send to the partition and wait for it
+                    let timer: metrics::ScopedTimerGuard<'_> =
+                        metrics.send_time[partition].timer();
+                    // if there is still a receiver, send to it
+                    if let Some((tx, reservation)) = output_channels.get_mut(&partition) {
+                        reservation.lock().try_grow(size)?;
+
+                        if tx.send(Some(Ok(batch))).await.is_err() {
+                            // If the other end has hung up, it was an early shutdown (e.g. LIMIT)
+                            reservation.lock().shrink(size);
+                            output_channels.remove(&partition);
+                        }
                     }
+                    timer.done();
+                } else {
+                    let timer: metrics::ScopedTimerGuard<'_> =
+                        metrics.send_time[partition].timer();
+                    // if there is still a receiver, send to it
+                    if let Some((tx, reservation)) = output_channels.get_mut(&partition) {
+                        reservation.lock().try_grow(size)?;
+
+                        if tx.send(Some(Ok(batch))).await.is_err() {
+                            // If the other end has hung up, it was an early shutdown (e.g. LIMIT)
+                            reservation.lock().shrink(size);
+                            output_channels.remove(&partition);
+                        }
+                    }
+                    timer.done();
                 }
-                timer.done();
             }
 
             // If the input stream is endless, we may spin forever and